MB-62230 - Optimizing cluster selection in pre-filtered kNN

[metonymic-smokey]

Modifies the pre-filtering kNN path for IVF indexes to factor in the distribution of filtered hits in the index.

1. Filter the clusters which have at least one filter hit.
2. Rank the filtered clusters in decreasing order of proximity.
3. Pick either the closest 'nprobe' clusters or as many as needed to satisfy at least K hits, whichever comes first.
4. Search the selected clusters for the k nearest neighbours from among the filtered hits.

There is no change in the pre-filtered search path for Flat indexes.